🧑‍🔬 Xavier Martinet

Xavier Martinet's papers here sit where large language models meet automated reasoning. The work deals with transformer models and how to train and use them. One line asks how to build capable language models that stay cheap to run, using only openly available text. The other applies neural networks to finding formal mathematical proofs in systems such as Metamath and Lean.

The most cited is LLaMA: Open and Efficient Foundation Language Models (2023). It presented a family of foundation models of several sizes, trained on far more tokens than earlier scaling-law guidance advised, so that smaller models could perform well at deployment. Its training corpus came entirely from public sources, including cleaned web crawls, code, Wikipedia, books and scientific preprints. The earlier HyperTree Proof Search for Neural Theorem Proving (2022) described the Evariste prover. Evariste uses an AlphaZero-inspired search that expands whole proof hypertrees rather than single paths. A transformer proposes tactics, and a critic estimates which goals can be proved. Both networks keep learning online as the search runs.

From 2022 to 2023 the papers move from search-based theorem proving toward general-purpose pretraining, under the topics Reasoning and Language models & pretraining. Both projects came from research teams at Meta AI. Collaborators who appear across this work include Marie-Anne Lachaux, Thibaut Lavril, Timothée Lacroix and Aurelien Rodriguez, with Guillaume Lample on the theorem-proving paper and Hugo Touvron on the language-model work.

Xavier Martinet has 2 papers in our library, in 2022–2023, most of them on Language models & pretraining, Reasoning; often with Aurelien Rodriguez, Marie-Anne Lachaux, Thibaut Lavril.